SUMMARY: Performance issues with multi processors

From: Jeff Kennedy (jkennedy@amcc.com)
Date: Thu May 18 2000 - 12:06:15 CDT


Thanks go to:

Bill Hebert Mark Bergman
Marc Johnson Buddy Lumpkin
Bismark Espinoza Mike Evans
Andrew Sit David Mitchell
Ken Robson Russ Poffenberger

The above suggested several things. One was www.spec.org for standard
benchmarks. Unfortunately we have different requirements than standard
benchmarks adn these conflict with what the engineer managed to produce
anyway. Another was an issue with context switching. I am unsure of
exactly what that is but I plan to find out. Mutex was also suggested
but after looking at it's man page I realized that I have no business
there. Didn't understand half the stuff there and it looks like it's
for the developer side of things anyway (of which I am most certainly
not). It was also suggested that if the job/process was not
multi-threaded there would be issues with this as well.

Special thanks go to:

Casper Dik Will Robinson

Will mentioned he had seen a similar problem on HP and it was the result
of processes moving around from cpu to cpu (context switching?) and the
overhead and state saving that was associated with all this movement.
His solution was a patch from HP, no such animal here. However, Casper
comes through again. Pbind was his suggestion and I will be looking at
this to solve this issue. This will bind certain processes to certain
cpu's, eliminating the overhead of cpu switching.

We will be running these jobs again in the near future.

Thanks again to all of the above.

~JK

Jeff Kennedy wrote:
>
> I have an engineer that is telling me they did some benchmarks on the
> Ultra 80 and 420. One job was run on these machines with one cpu, then
> 1 job was added with an additional cpu up to 4. In the end they had 4
> of the same jobs running on a 4 cpu machine.
>
> They said they saw a 5% + decrease in performance for each job each time
> a cpu/job was added. These jobs are not memory intensive (around 30MB
> each) and do not use disk (simulation jobs). Their theory is that it
> has something to do with the backplane, but at 1.6 GB/sec I am having a
> hard time believing that.
>
> Anyone have any background in this sort of thing or thoughts as to why
> this might happen? I haven't done these myself so I'm running on their
> word at this point.
>

-- 

=================== Jeff Kennedy UNIX Administrator AMCC jkennedy@amcc.com



This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:14:08 CDT